Re: effect of diseases on length of stay

Posted by Richard Ristow on Jan 18, 2007; 11:31pm
URL: http://spssx-discussion.165.s1.nabble.com/no-subject-tp1073249p1073263.html

Well, I'll start, but others can add more than I can.

At 11:59 AM 1/16/2007, Angshu Bhowmik wrote:

>I have data on lengths of hospital stay (LOS, in days) in about 100
>patients with Disease A (always present i.e. 1). There are various
>co-morbidities: Disease B, Disease C etc which may be present or
>absent. There are in total 18 co-morbidities, but they could be
>grouped into 5 groups if that makes it easier to perform a more
>sensible analysis.

That grouping is the only way to do it. It gives you a mean of 20
patients per group, which is reasonable. Keep all 18 co-morbidities,
and you have a mean of 100/18=5.5, which isn't going to be enough.

Then, you're comparing means between groups: that's a one-way analysis
of variance. In SPSS, command MEANS is a good place to start. The
syntax is easy, and it allows a lot of descriptive statistics by cell.
For descriptives, I'd select COUNT, MEAN, STDDEV, MEDIAN, MIN and MAX;
with /STATISTICS=ANOVA. From the menus: Analyze>Compare means>means.
Don't do a test for linearity; it's not meaningful, for you.

Moving on,

. If the F-test says the groups differ, you'll likely want to know
which groups have significantly higher or lower means, than which
others. That's called multiple comparison analysis, and is available in
command ONEWAY (Menu Analyze>Compare means>ANOVA). Select "post hoc",
and pick a test - try BONFERRONI first, but that's something others on
the list will know more about than I do.

. I don't know the shape of your length-of-stay distributions: Do they
cluster around a value? Or are there a lot of short stays, and a small
proportion of much longer ones? When you run your ANOVA, look for cell
means substantially larger than the medians; that can point to the
latter. If it's the case, you may want to use a non-parametric ANOVA.
Or, I'd seriously consider log-transforming the data. There may be
other opinions on that, though. A lot of people, including me, are very
cautious about transforming data to make it look 'nice'. That said, the
log transform still feels like something to try, if the distributions
show lengthy tails.

Now, is that enough to get you going?

-Best of luck,
  Richard